The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Supplementary Material to Working Paper 08-07 Sectoral vs. Aggregate Shocks: A Structural Factor Analysis of Industrial Production Andrew T. Foerster Department of Economics, Duke University Pierre-Daniel G. Sarte Research Department, Federal Reserve Bank of Richmond Mark W. Watson Department of Economics, Princeton University 1 1 The Model max Et ∞ X βt t=0 subject to Yjt = Cjt + N X j=1 N X i=1 and Yjt = à 1−σ −1 Cjt − ψLjt 1−σ ! Mjit + Kjt+1 − (1 − δ)Kjt α Ajt Kjtj N Y γ i=1 The first-order necessary conditions are: S 1−αj − N i=1 γ ij Mijtij Ljt . −σ = λjt , Cjt : Cjt à ! N X Yjt Ljt : ψ = λjt 1 − αj − γ ij . Ljt i=1 Combining these two equations gives −σ Yjt ψ = Cjt Ljt à 1 − αj − Mijt : λit = λjt γ ij or −σ γ ij Cit−σ = Cjt Kjt+1 : 2 −σ Cjt N X γ ij i=1 ! . Yjt , Mijt Yjt . Mijt ∙ µ ¶¸ Yjt+1 −σ = βEt Cjt+1 αj +1−δ Kjt+1 Dynamics of the System The dynamics are described by a set of 4N + N 2 equations in 4N + N 2 unknowns. When N = 117, this amounts to 14157 equations, but preliminary algebraic manipulations help keep the system tractable. à ! N X −σ ψLjt = Cjt 1 − αj − γ ij Yjt i=1 −σ Cit−σ = Cjt γ ij 2 Yjt Mijt −σ Cjt ∙ µ ¶¸ Yjt+1 −σ = βEt Cjt+1 αj +1−δ Kjt+1 Yjt = Cjt + N X i=1 and Yjt = Mjit + Kjt+1 − (1 − δ)Kjt α Ajt Kjtj N Y i=1 3 S γ ij 1−αj − N i=1 γ ij Mijt Ljt . Log-linearized Equations The “hat” notation stands for percent deviation from steady state. bjt + Ybjt bjt = −σ C L bit = −σ C bjt + Ybjt − M cijt −σ C eYbjt+1 − β eK bjt = −σEt C b jt+1 bjt+1 + β −σ C e = 1 − β + βδ. where β bjt + SK K b jt+1 − (1 − δ)SK K b jt + Ybjt = SCj C j j bjt + αj K b jt + Ybjt = A N X i=1 à cijt + 1 − αj − γ ij M N X i=1 cjit SMji M N X γ ij i=1 ! bjt . L b1t , ..., C bNt ], etc... and mt = [M c11t , ..., M c1Nt , M c21t , ..., M cNN t ]. The log-linearized Let ct = [C equations can be written in matrix form as follows: lt = −σct + yt , (1) mt = My yt + Mc ct (2) where My = 1N×1 ⊗ I and Mc = σ(I ⊗ 1N×1 ) − σ(1N ×1 ⊗ I), e t yt+1 − βk e t+1 −σct = −σEt ct+1 + βE (3) yt = Sc ct + Sm mt + Sk kt+1 − Sk (1 − δ)kt , (4) where ⎡ ⎢ Sc = ⎣ ⎡ ⎤ SC1 ... SCN ⎢ ⎥ ⎦ , etc... and Sm = ⎣ 3 SM11 SM12 ... 0 0 0 0 ... SMN,N −1 SMN N ⎤ ⎥ ⎦, e t + Φlt , yt = at + αd kt + Γm where eN ×N 2 Γ ⎡ γ 11 0 ... γ 21 ⎢ ⎢ 0 γ 12 ... 0 ⎢ =⎢ γ 13 ⎢ ⎢ ... ⎣ and γ 1N ⎡ P ⎢ ⎢ Φ = I − αd − ⎢ ⎢ ⎣ 4 0 γ 22 i γ i1 (5) ... γ N1 0 ... ... γ N2 γ 23 γ N3 ... ... 0 ... γ 2N γ NN P i ⎤ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎦ ⎤ γ i2 ... P i System Reduction γ iN ⎥ ⎥ ⎥ ⎥ ⎦ Use equation (1), (2) and (5) to obtain e e y − Φ]yt = at + αd kt + [ΓM − Φσ]ct [I − ΓM | c{z } {z } | Ωyc αd or, equivalently −1 yt = α−1 d at + kt + αd Ωyc ct . Note that Ωyc = αd (I − (I − Γ0 )α−1 d ) when σ = 1. Substituting this equation in equation (3) gives e −1 at+1 + kt+1 + α−1 Ωyc ct+1 ] − βk e t+1 −σct = −σct+1 + β[α d d or e −1 Ωyc ]ct+1 + βα e −1 at+1 . −σct = [−σI + βα d d Use the resource constraint (4) to obtain yt = Sc ct + Sm [My yt + Mc ct ] + Sk kt+1 − Sk (1 − δ)kt or (I − Sm My )yt = (Sc + Sm Mc )ct + Sk kt+1 − Sk (1 − δ)kt which gives −1 (I − Sm My )[α−1 d at + kt + αd Ωyc ct ] = (Sc + Sm Mc )ct + Sk kt+1 − Sk (1 − δ)kt 4 (7) or finally Sk kt+1 = [(I − Sm My )α−1 d Ωyc − (Sc + Sm Mc )]ct + [Sk (1 − δ) + (I − Sm My )]kt (8) +(I − Sm My )α−1 d at We can write equations (7) and (8) as: # " # " e −1 Ωyc 0 c −σI + βα t+1 d Et 0 Sk kt+1 # #" " ct −σI 0 + = kt (I − Sm My )α−1 d Ωyc − (Sc + Sm Mc ) Sk (1 − δ) + (I − Sm My ) " # # " e −1 0 −βα d at + Et (at+1 ) −1 (I − Sm My )αd 0 (9) At this stage, the dynamics of the system can be solved using standard linear rational expectations toolkits such as Blanchard and Kahn (1980), King, Plosser, Rebelo (1988), and Klein (2000). The results presented in the text are based on King and Watson (2002). To use these methods, however, one must first obtain the steady state of the system. In this model, this can be achieved analytically. 5 Finding the Steady State Analytically The steady state solution only requires inverting N ×N matrices. The steady state equations for labor, materials, and capital are respectively ψLj = λj (1 − αj − M1j = N X γ ij )Yj i=1 λj γ Yj λi ij 1 Kj = αj [ − 1 + δ]−1 Yj . β | {z } φKj Now take the logs of these equations to obtain (small letters denote logs) lj = − ln ψ + ln λj + ln(1 − αj − N X γ ij ) + yj , i=1 mij = ln λj − ln λi + ln γ ij + yj 5 kj = ln φKj + yj . (12) The log steady state equations can be written in matrix form to summarize the entire system. Let l = [l1 , ..., lN ], etc... and m = [m11 , m12 , ...m1N , m21 , ...mNN ]. Then, we have that e + ln λ + y, l = − ln ψ + ln Φ where ⎡ ⎤ P ln(1 − α1 − i γ i1 ) ⎥ e=⎢ ln Φ ... ⎣ ⎦. P ln(1 − αN − i γ iN ) Similarly, the equation for steady state materials can be expressed as, m = Mλ ln λ + My y + vec(ln Γ0 ), where Mλ = 1N×1 ⊗ I − I ⊗ 1N×1 , and My = 1N×1 ⊗ I. Finally, we have that k = ln φK + y, where ⎡ ⎤ ln φK1 ⎢ ⎥ ln φK = ⎣ ... ⎦ ln φKN The log of production in all sectors can be expressed as e + Φl, y = a + αd k + Γm e is defined as above. Making the appropriate substitutions yields where Γ e λ ln λ + My y + vec(ln Γ0 )] + Φ(− ln ψ + ln Φ e + ln λ + y) y = a + αd (ln φK + y) + Γ[M or, equivalently, e e e − Φ ln ψ + (ΓM e λ + Φ) ln λ. [I − αd − ΓM − Φ]y = a + αd ln φK + Γvec(ln Γ0 ) + Φ ln Φ {z y } | =0 It follows that we can solve for (shadow) prices in the steady state in closed form, e e − Φ ln ψ], e λ + Φ)−1 [a + αd ln φK + Γvec(ln Γ0 ) + Φ ln Φ ln λ = −(ΓM and λ = eln λ . 6 To solve for the vector Y , write the resource constraints as 1 λ− σ + δφdK Y + Mr Y = Y where φdK ⎡ ⎢ ⎢ =⎢ ⎢ ⎣ ⎤ φK1 ... φKN −1 ΦKN ⎡ ⎥ ⎢ ⎥ ⎢ ⎥ , and Mr = ⎢ ⎥ ⎢ ⎦ ⎣ γ 11 γ 12 λλ21 γ 21 λλ12 γ 22 ... λ1 γ N 1 λN γ N2 λλN2 ... γ 1N λλN1 ... γ 2N λλN2 ... ... γ NN ⎤ ⎥ ⎥ ⎥, ⎥ ⎦ and φdK is a diagonal matrix with φK on its diagonal. The solution for Y is then given by 1 Y = [I − δφdK − Mr ]−1 λ− σ . Solving for the remaining variables in the steady state is then straightforward. 6 Output from King and Watson (2002) programs The policy functions take the form (with 2 sectors as an example): ⎤⎡ ⎡ ⎤ ⎡ ⎤ k1t ... c1t ⎥⎢ ⎢ ⎥ ⎢ ⎥ ⎥ ⎢ k2t ⎥ ⎢ c2t ⎥ ⎢ ⎥⎢ ⎢ ⎥=⎢ ⎥ ⎢ k ⎥ ⎢ 1 0 0 0 ⎥⎢ δ ⎥, ⎦ ⎣ 1t ⎦ ⎣ 1t ⎦ ⎣ 0 1 0 0 k2t δ 2t " # " a1t ... = a2t | {z } 1 0 0 1 xt and ⎡ ⎢ ⎢ ⎢ ⎢ ⎣ k1t+1 k2t+1 δ 1t+1 δ 2t+1 k1t k2t δ 1t δ 2t #⎢ ⎢ ⎢ ⎢ ⎣ ⎡ ⎤ ⎡ #⎢ ⎥ " ⎢ ⎥ M M k a ⎢ ⎥= ⎢ ⎥ 0 I ⎣ ⎦ k1t k2t δ 1t δ 2t ⎤ ⎤ ⎥ ⎥ ⎥, ⎥ ⎦ ⎥ ⎥ ⎥ + Hεt . ⎥ ⎦ More generally, we can write these equations as # " # " #" Πck Πca ct kt = , kt I 0 δt | {z } st 7 xt = and, 7 " kt+1 δ t+1 # = " h Q Mk Ma 0 I i st #" kt δt # + Hεt . Obtaining the Filtering Matrices Since we assume that the logarithm of sectoral productivity follows a random walk, Q = I in the procedure governing the driving process (i.e. drp.gss) of King and Watson (2002). Then, we have that kt+1 = Mk kt + Ma at while ct = Πck kt + Πca at . Recall that −1 yt = α−1 d at + kt + αd Ωyc ct . Therefore, −1 yt = α−1 d at + kt + αd Ωyc [Πck kt + Πca at ] [I + Ωyc Πca ]at + [I + α−1 = α−1 d Ωyc Πck ]kt {z } {z } |d | Πa so that Πk −1 kt = Π−1 k yt − Πk Πa at . Using these equations, we have that yt+1 = Πk kt+1 + Πa at+1 = Πk (Mk kt + Ma at ) + Πa at+1 −1 = Πk Mk (Π−1 k yt − Πk Πa at ) + Πk Ma at + Πa at+1 or yt+1 = Πk Mk Π−1 y + Π (M − M Π−1 Π )a + Πa at+1 . | {z k } t | k a {z k k a} t % Ξ Under the assumptions made in the paper regarding the process for at , it follows that ∆yt+1 = %∆yt + Ξεt + Πa εt+1 , 8 so that the filtering is carried out according to −1 −1 εt+1 = Π−1 a ∆yt+1 − Πa %∆yt − Πa Ξεt . where ε0 is set to zero.1 Let η t+1 = Ξεt + Πa εt+1 , Then, if var(εt ) = I, Σηη = ΞΞ0 + Πa Π0a . References [1] Blanchard, O. and C. Kahn (1980), “The solution of linear difference models under rational expectations, Econometrica, 48:1305-1311. [2] King, R. G., Plosser, C. I, and S. T. Rebelo (1988), “Production, growth, and business cycles, technical appendix,” manuscript, University of Rochester. [3] King, R. G., and M. W. Watson (2002), “System reduction and solution algorithms for singular linear difference systems under rational expectations,” 20:57-86. [4] Klein, P. (2000), “Using the generalized Schur form to solve a multivariate linear rational expectations model,” Journal of Economic Dynamics and Control, 24:1405-1423. 1 For the various calibrations presented in the text, the eigenvalues of Π−1 a Ξ have modulus less than one. 9